Add Anthropic tool search bridge for deferred tool definitions

[phfarath]

Summary

First slice of #766: a provider-bridged tool-search abstraction, implemented for Anthropic's native mechanism.

With many MCP tools, every definition is serialized into each request, costing tokens and degrading selection accuracy past ~30-50 tools. This adds a RequestParams.tool_search policy (off/auto/always, default off) and bridges it to Anthropic tool search: regular tool definitions are marked defer_loading: true and a server-side BM25 tool_search tool is appended, so the model retrieves definitions on demand. Search runs server-side; the agent only receives the final tool_use for the real tool, so the execution loop is unchanged (name resolution already tolerates definitions that were never loaded, and *_tool_result blocks — including tool_search_tool_result — are already persisted/replayed by the existing server-tool channel machinery).

Changes

• RequestParams.tool_search: Literal[off,auto,always] (default off), excluded from provider payloads via BASE_EXCLUDE_FIELDS
• AnthropicLLM: _defer_tool_definitions() marks tools defer_loading and appends the tool_search_tool_bm25_20251119 tool param; _resolve_tool_search_active() resolves the policy
• Disabled on anthropic-vertex (beta allowlist), suppressed for structured tool_use requests, and disabled when a sampling tool_choice pins specific tools
• auto defers above 16 tools
• Docs: Anthropic provider page + generated request-params reference

Testing

• Unit tests: tests/unit/fast_agent/llm/provider/anthropic/test_tool_search.py (policy resolution, defer marking, structured suppression, Vertex gating, threshold)
• Local payload E2E: real app/agent/tool-runner pipeline with network stubbed at the SDK boundary — confirmed defer_loading: true + search tool present with always, absent by default
• Live API smoke (haiku): API accepted the payload; model searched server-side, called deferred tools by name; turn-2 history replay with tool_search_result blocks worked; control run unchanged
• format --check, lint, typecheck clean on changed files; llm suite 1693 passed (6 pre-existing failures reproduced on clean baseline)

Follow-ups (out of scope here)

• Bridge defer_loading for provider-managed remote MCP on Anthropic (currently OpenAI Responses only)
• Generic client-side search tool fallback for providers without native support

Addresses #766.

---